Gemma 4 31B
Google's largest dense Gemma 4 model, released with open weights under Apache 2.0 in April 2026 and also offered in Google AI Studio and the Gemini API. It is far behind frontier models but is a popular choice for running AI locally and privately.
GoogleLanguage modelReleased April 2026Data checked October 18, 2026
This model is not in the current people's rankings: it is either new and still collecting votes, or ranked below the top 12. See the AI leaderboard for the models people rate highest.
On paper
Fixed-test rankings from the benchmarks tab: Speed #23, Reasoning #24.
Strengths and weaknesses
- Apache 2.0 license and about 9.1 million downloads on Hugging Face
- Image input and a 256K-token context window
- A higher LMArena text score than Grok 4.7 and GPT-6 Luna despite its small size
- Lowest Artificial Analysis Intelligence Index in this list (14.7)
- Weak at agentic coding (0% on Terminal-Bench 4.0 as run by Artificial Analysis, and last among these models on the LMArena Code Arena)
Best for
- Running AI on your own computer or server
- Private, offline use
- Learning and experimenting
Every score we have
Each bar compares this model with the best result any model on this page has on the same test.
- LMArena text arena (human votes)1,452.2Arena score 1452.2 (95% CI 1444.7 to 1459.6); 6,147 votes; entry "gemma-4-31b"; position 76 of 416 on the full board
- LMArena creative writing1,418.8Arena score 1418.8 (95% CI 1400.0 to 1437.6); 974 votes; entry "gemma-4-31b"; position 86 of 414 on the full board
- LMArena WebDev (coding)1,367.4Arena score 1367.4 (95% CI 1361.1 to 1373.6); 14,490 votes; entry "Gemma 4 31B"; position 106 of 144 on the full board
- LMArena vision1,262.1Arena score 1262.1 (95% CI 1256.3 to 1267.9); 39,669 votes; entry "gemma-4-31b"; position 44 of 165 on the full board
- Artificial Analysis Intelligence Index14.714.7 on the Artificial Analysis Intelligence Index v4.3.2 (AA shows 15), setting "Gemma 4 31B"
- Humanity's Last Exam (Artificial Analysis)23.6%23.6% on Humanity's Last Exam (AA, 2,158 text-only questions), setting "Gemma 4 31B"
- GPQA Diamond, graduate-level science85.7%85.7% on GPQA Diamond (AA; no longer part of the AA index since v4.2), setting "Gemma 4 31B"
- Terminal-Bench 4.0 (Artificial Analysis run)0.0%0.0% on Terminal-Bench 4.0 as run by AA (mini-swe-agent harness), setting "Gemma 4 31B"
- Long-context reasoning (AA-LCR)69.7%69.7% on AA-LCR (long-context reasoning), setting "Gemma 4 31B"
- MMMU-Pro, image understanding73.4%73.4% on MMMU-Pro (image understanding, AA), setting "Gemma 4 31B"
- IFBench, following instructions75.6%75.6% on IFBench (instruction following, AA), setting "Gemma 4 31B"
- AA-Omniscience, accuracy minus hallucination-47.9-47.9 on the AA-Omniscience Index (knowledge accuracy minus hallucination, range -100 to 100), setting "Gemma 4 31B"
- Output speed, tokens per second35 t/s35 output tokens per second (AA median), row "Gemma 4 31B"
- Cost per Artificial Analysis test task$0.00$0.00 to run one AA Intelligence Index task, row "Gemma 4 31B"
Compare with
- Claude Opus 5.5Anthropic's flagship for long-running agentic coding and knowledge work, and the model Anthropic suggests starting with for most jobs.
- Claude Sonnet 5.5Anthropic's mid-priced model, which Anthropic describes as its best mix of speed and intelligence.
- Claude Fable 5.1Anthropic's top tier for the hardest reasoning and long-horizon agent work, which Anthropic recommends when Opus 5.5 still falls short.
- Claude Haiku 5.5Anthropic's smallest and fastest current model, built for high-volume, latency-sensitive jobs like classification, extraction and routing.